The semantic segmentation task is to classify the object category for each pixel in the image, so its complexity is higher than that of general object detection task. Currently, the challenge of semantic segmentation lies in how to perform image segmentation at a high frame rate and with high accuracy, as well as how to deploy it on low-computing-power platforms. Therefore, this paper researches a lightweight BiSeNet semantic segmentation network that combines YOLOv5 feature fusion network. This network is based on BiSeNet (bilateral segmentation network) and introduces MobileNet v2 as the backbone to reduce model parameters. Additionally, the YOLOv5 feature fusion layer is embedded to obtain enhanced multi-scale feature maps and improve the model's segmentation accuracy. Experiments show that the Mean Intersection over Union (MIoU) of the prosed method on the CamVid dataset is 1.83% higher than the one before improvement, and the model parameters are reduced by 71.6%.